CiteWeave for DeepSeek Harness

Web search with sourced images for DeepSeek Harness.

简体中文 · Backend · Reference · Releases

CiteWeave retrieves web evidence, evaluates candidate images, and supplies source links and illustrations to Harness. The configured DeepSeek model writes the answer, placing images beside the relevant explanation.

This repository contains the Harness plugin and its Python backend. The backend runs Qwen3.5-9B through vLLM, with CPU embeddings and reranking. Celery, Redis, and PostgreSQL handle retrieval jobs. On shared GPUs, the model starts on demand and releases memory after an idle period.

Examples

Geology — Grand Canyon rock layers with a National Park Service diagram.

Biology — Monarch identification with a photograph of its wing markings.

Engineering — A suspension bridge's load path, illustrated with the Golden Gate Bridge.

Image credits

Installation

Backend

Deploy the service from backend/. The reference configuration uses Linux, Python 3.12, Docker, an RTX 3060 12 GB, and 32 GB RAM.

git clone https://github.com/wavachao/dsh-citeweave.git
cd dsh-citeweave/backend

Follow the deployment guide to install dependencies, download the models, and start the services.

Harness plugin

In Harness Desktop or Web, open Plugins → Add plugin and enter:

github:wavachao/dsh-citeweave#v0.5.1

Open Installed plugins → CiteWeave → Configure CiteWeave, enter the backend address and workspace access token, then save.

Setting	Default	Description
endpoint	http://127.0.0.1:8765	Backend address; also accepts CITEWEAVE_ENDPOINT.
accessToken	Empty	Workspace token; also accepts CITEWEAVE_ACCESS_TOKEN.
timeoutMs	300000	Tool timeout in milliseconds.
maxImages	3	Maximum images per search, from 1 to 8.

The token field preserves the saved value when left blank. Command-line installation and environment configuration are covered in the reference.

Usage

Explain how Eratosthenes measured Earth's circumference. Include a diagram and cite its source.

Compare suspension bridges and cable-stayed bridges, with photographs showing their structural differences.

Harness calls citeweave_search, incorporates the evidence, and embeds images through its attachment store. The tool supports automatic search, up to five source URLs, image verification, and cache refresh. See the tool reference.

If the backend is unavailable, Harness continues answering without that tool result. Retrieval coverage and image availability depend on the source websites.

Evaluation

See benchmarks for comparison scripts, sample answers, and results.

Project structure

Path	Contents
index.js, client.js	Harness tool and configuration interface
backend/src/citeweave/	Retrieval API, worker, ranking, and model lifecycle
backend/	Python package, deployment configuration, and tests
benchmarks/	Harness comparison and measurements

The plugin targets Harness 0.2.1-rc.2 and Node.js 24+. Plugin and backend source share versioned releases; the installation archive contains the Harness plugin. Backend deployments use the matching repository tag.

Contributing

See CONTRIBUTING.md for development and tests. Report bugs through GitHub Issues, and vulnerabilities through SECURITY.md.

License

MIT. Based on CiteWeave. Model and third-party resource licenses are listed in backend/THIRD_PARTY.md; retrieved images retain their original licenses.